Abstract Breast cancer remains a major cause of cancer-related mortality worldwide. Despite advances in treatment, toxicity and drug resistance arouse a need for safer therapeutics, meaning that researchers must develop innovative computational strategies to accelerate drug discovery. In this study, we developed a predictive machine learning model using cytotoxicity data from the MCF-7 breast cancer cell line, curated from the ChEMBL database. The raw dataset contained 55,210 compounds, which, after rigorous filtering for assays employing the MTT method with 72-hour exposure and nanomolar (nM) activity values, yielded 3,753 compounds. These were subsequently preprocessed and feature engineered to generate molecular descriptors and fingerprints as independent variables (X). Experimental IC₅₀ values were converted into binary activity classes, with compounds ≤ 1000 nM labelled active and those > 1000 nM inactive, serving as the dependent variable (y). Three classifiers, k-Nearest Neighbours (KNN), Extra-Trees, and LightGBM, were evaluated on a test set, achieving accuracies of 0.9294, 0.9348, and 0.9334, respectively. Consensus predictions from these models were used to screen a natural product database, followed by drug-likeness evaluation via Lipinski’s RO5, similarity-based pharmacophore defined by the molecular weight and spatial arrangement of key functional groups, and then molecular docking against ERα. The top-ranked compounds, 12-hydroxy-bullatacin A, rollitacin, and gigantecin, exhibited strong binding affinities of − 11.6, − 11.0, and − 10.6 kcal/mol, outperforming the reference drug 4-hydroxytamoxifen (− 9.6 kcal/mol). This integrative approach, combining biologically curated cytotoxicity data, consensus machine learning, and structure-based screening, provides a reliable framework for repurposing and prioritising natural compounds with potential anti-breast cancer activity.
Johnson et al. (Tue,) studied this question.